Rafael Peñaloza Nyssen is an associate professor in the Department of Computer Science, Systems and Communication at the University of Milan-Bicocca (for those who are curious, he also has a blog where, when he has the time and the right inspiration, he discusses topics related to Artificial Intelligence). He has long been involved in Artificial Intelligence, particularly in Knowledge Representation and Reasoning, including in situations where information and knowledge are incomplete or imperfect. A few months ago, in collaboration with other researchers from the Department and as part of the ReGAInS project, he launched a series of seminars entitled the Neuro-Symbolic Seminar (NeSS).

Q: Rafael, for those who aren’t particularly familiar with machine learning – and perhaps aren’t even very familiar with AI – what is the Neuro-Symbolic Seminar (NeSS) series about?

Historically, AI has been divided into two camps: connectionist approaches, which seek to uncover hidden patterns within a dataset that enable an intelligent system to predict the behaviour of new data; and symbolic approaches, which instead start from structured knowledge to arrive at new conclusions. For many years, the prevailing debate centred on which of these two approaches was best suited to achieving true ‘Artificial Intelligence’. We now realise that both are essential. The aim of NeSS is to promote new ideas and accelerate the development and adoption of neuro-symbolic AI (which integrates connectionist and symbolic ideas) towards a more reliable, accurate and secure form of AI.

Q: For those who are somewhat familiar with AI, at least as it has been discussed in recent years. Why talk about symbolic approaches at a time when machine learning seems to be the dominant approach?

Machine learning has proved crucial for solving many problems where a formal, structured definition is impossible or impractical. We cannot easily describe what distinguishes an image of a dog from that of a cat, recognise a specific fingerprint, or identify a voice command. Yet neural networks are able to learn the implicit properties of this information and carry out these tasks. However, the strength of these methods is also their greatest weakness: by design, they are imprecise and not interpretable; furthermore, they require vast amounts of training data. On the other hand, it makes no sense to discard thousands of years of knowledge acquired by humankind; knowledge which, moreover, often goes beyond mere data. To send a rocket to Mars, one needs to know the laws of celestial mechanics and the varying gravitational pull. These are precise and well-known. As mentioned earlier, we are not suggesting that neural methods are useless. Rather, we want to enhance them with specific, precise and formal knowledge, so that they can reason and make rational decisions.

Q: Who is the main target audience for the seminars? How is the initiative progressing?

So far, we have held four seminars led by world-class scientists based in the USA, the UK, Belgium and Italy. These seminars were primarily aimed at researchers specialising in neural or symbolic methods who wish to familiarise themselves with the neuro-symbolic literature. In future, we hope to broaden our scope to include more accessible seminars for a less specialised audience.

For anyone interested, the seminars have been recorded and can be watched again on this page.